Mushroom Language Project: Exploring Spontaneous Language Evolution of Neural Networks in Open Environments

The Mushroom Language Project is a modern implementation based on Cangelosi's 1998 classic experiment, aiming to explore the process by which neural network populations spontaneously form communication languages in non-fragmented environments through open-ended neuroevolution and endogenous selection mechanisms. The core of the project is to study the emergence of language in artificial intelligence systems, bridging classic experiments with modern computational methods.

In 1998, Angelo Cangelosi's classic experiment first demonstrated the possibility of artificial neural network populations spontaneously developing primitive communication systems under evolutionary pressure. In the experiment, senders could perceive the location of food but could not move, while receivers could move but could not see the food; the two communicated through a discrete signal channel. After multiple generations of evolution, they formed a primitive language linking food locations to signals, revealing that language is an adaptive coordination mechanism under evolutionary pressure.

The core innovation of the project lies in the design of non-fragmented environments: creating a continuously running world where agents can be born, interact, reproduce, and die—closer to real evolutionary processes, allowing complex social structures and cultural inheritance. Additionally, it adopts open-ended neuroevolution methods, allowing dynamic changes in network topology (connections, number of layers, activation functions), enabling the discovery of unexpected solutions.

The project introduces an endogenous selection mechanism, different from traditional external fitness functions: the survival and reproduction of agents are determined by interaction dynamics (such as resource competition, sexual selection, kin selection), and evolutionary pressure itself becomes a product of co-evolution. This mechanism also raises thoughts on AI safety: when goals emerge from within, how to ensure they align with human values?

In terms of technical implementation, the project may use the NEAT/HyperNEAT framework to optimize network weights and topology; non-fragmented environments handle asynchronous events and state persistence through agent-based simulation frameworks; signal channel design may explore continuous, multi-dimensional, or learning-based protocols, affecting language complexity. Specific details have not yet been fully disclosed.

This research is of great significance to AGI and cognitive science: unlike LLMs (which compress and reorganize human language), it explores the possibility of agents developing communication systems from scratch, potentially providing AGI with autonomous communication capabilities; at the same time, it provides a computational model for the origin of human language, bridging computational science and cognitive science.

The project faces challenges: non-fragmented environments increase computational complexity and analysis difficulty; the identification and quantification of language emergence remain unresolved; the growth in the number of agents leads to an exponential rise in interaction complexity. Looking ahead, it can integrate deep learning to handle high-dimensional perception, collaborate with multi-agent RL, and even realize cultural evolution in AI.

The Mushroom Language Project is not just code and experiments; it is a window to understanding the essence of intelligence. It reminds us that language may be a natural product of the information processing needs of complex systems. For researchers and enthusiasts, this project connects the classic and the modern, combining theoretical depth with engineering challenges, and is worthy of attention and exploration.